Google Releases Gemma Scope 2 to Deepen Understanding of LLM Behavior

"Gemma Scope 2 is a suite of tools designed to interpret the behavior of Gemini 3 models, enabling researchers to analyze emergent model behaviors, audit and debug AI agents, and devise mitigation strategies against security issues like jailbreaks, hallucinations and sycophancy. Interpretability research aims to understand the internal workings and learned algorithms of AI models. As AI becomes increasingly more capable and complex, interpretability is crucial for building AI that is safe and reliable."

"Google describes Gemma Scope as a microscope for its LLMs. It combines sparse autoencoders (SAEs) and transcoders to let researchers inspect a model's internal representation, examine what it "thinks" and understand how those internal states shape its behavior. One key use case is inspecting discrepancies between a model's output and its internal state, which Google says could help surface safety risks."

"Gemma Scope 2 extends the original Gemma Scope, which targeted the Gemma 2 family, in several ways. Most notably, it retrained its SAEs and transcoders across every layer of Gemma 3 models, including skip-transcoders and cross-layer transcoders, which are designed to make multi-step computations and distributed algorithms easier to interpret. Increasing the number of layers, Google explains, directly increases compute and memory requirements, which required to design specialized sparse kernels to keep complexity scaling linearly with the number of layers."

Gemma Scope 2 provides tools to interpret Gemini 3 model behavior, enabling analysis of emergent behaviors, auditing and debugging of AI agents, and development of mitigation strategies for security issues such as jailbreaks, hallucinations, and sycophancy. The system combines sparse autoencoders (SAEs) and transcoders to inspect internal representations, reveal what the model "thinks", and trace how internal states shape outputs. SAEs and transcoders were retrained across every layer of Gemma 3, including skip and cross-layer transcoders to aid interpretation of multi-step computations. Specialized sparse kernels keep compute and memory scaling linear, and advanced training improves concept identification and addresses prior flaws. Additional tools focus on chatbot analysis, refusal mechanisms, and chain-of-thought faithfulness.

Read at InfoQ